A New Scheme on Privacy Preserving Association Rule Mining

1
A New Scheme on Privacy Preserving Association
Rule Mining

• Presentation in PKDD 2004

2
Privacy Concern

• Survey on net users' attitudes about online
  privacy shows that there are
• 7 privacy fundamentalists
• 56 pragmatic majority
• 27 marginally concerned
• Reference L. F. Cranor, J. Reagle, and M. S.
  Ackerman, Beyond ConcernUnderstanding Net
  Users' Attitudes about Online Privacy, ATT
  Labs-Research, Technical Report TR 99.4.3, 1999

3
Privacy Preserving Data Mining

• B2B (Business to Business)

• B2C (Business to Customer)

4
Association Rule Mining

• Privacy individual transactions
• Goal Identify frequent itemsets accurately while
  keeping private information confidential

5
Basic Notions

• We have

6
Privacy Preserving ASM

• Previous Approach Randomization

• Given transaction t, randomly choose
• Reference A. Evfimievski, R. Srikant, R.
  Agrawal, and J. Gehrke, Privacy Preserving
  Mining of Association Rules, in Proc. ACM SIGKDD
  2002, pp. 217-228.

7
Previous Approaches

• One-way communication from data providers to the
  data miner
• Transaction Invariant
• Item Invariant

8
Problems

• Need a better tradeoff between privacy and
  accuracy
• Failure on transactions with more than 10 items
• Inaccurate support for itemsets with more than 3
  items

9
Our New Scheme

• An Algebraic Approach

Transparent
Perturbation Guidance
10
Our New Scheme (cont.)

• Infrastructure

11
System Assumption

• Semi-honest behavior model Illegal data miners
  are honest but curious
• follow the protocol strictly (honest)
• attempt to learn private information from
  received messages (curious)
• Reference O. Goldreich, Secure Multi-Party
  Computation. Working Draft, 2002.

12
Perturbation Guidance Component

• Vk first k eigenvectors of
• Data miner side

13
Data Perturbation Component

• Truncated eigenvectors transformation
• Data provider side

14
Key Observations

• Observation 1
• Elements of are the support of
  2-itemsets
• By preserving the value of A, we can identify
  frequent itemsets with size 1, 2 and above.

15
Key Observations

• Observation 2
• Thus, Vk can help preserving A.

16
Negotiation

• Change of k with number of received transactions

17
Negotiation

• Change of accuracy with µ

18
Negotiation

• Perturbation level k

19
Net Users Attitude toward Privacy

• By negotiation, we can satisfy them all

• Survey on net users' attitudes about
• online privacy shows that there are
• 7 privacy fundamentalists
• 56 pragmatic majority
• 27 marginally concerned

20
Accuracy

• Maximum of false positive and false negative

21
Privacy

• What we care and what we do not care
• Frequent itemsets (wanted by data miner)
• Infrequent itemsets (unwanted by data miner and
  hence private)
• Metric
• With some approximation, we have

22
Simulation Results
23
Performance on Real Datasets
24
Future Work

• Algebraic approach on classification and
  clustering
• Combination of data transformation and query
  restriction approaches

25
Thank you

• Questions
• email nzhang_at_cs.tamu.edu